Team reports 99.9% Sudoku-Extreme accuracy with fewer than 250,000 parameters
The team says adding persistent memory and removing timestep conditioning turned a denoising model into an iterative solver.
TLDR
The team reports 99.9% accuracy on Sudoku-Extreme and 98.3% on Maze-Unique with a model containing fewer than 250,000 parameters. It describes the modified denoising model as an “anytime iterative solver.” A team member says retaining memory between steps “seems key,” drawing a parallel with classic Sudoku solvers using memory to track possible states.
Combined views
—
1 Source, first seen ago
